Novelty detection for breast cancer image classification
Paweł Cichosz, Dariusz Jagodziński, Mateusz Matysiewicz, Łukasz Neumann, Robert Marek Nowak, Rafał Okuniewski, Witold Oleszkiewicz · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2016
Using classification learning algorithms for medical applications may require not only refined model creation techniques and careful unbiased model evaluation, but also detecting the risk of misclassification at the time of model application. This is addressed by novelty detection, which identifies instances for which the training set is not sufficiently representative and for which it may be safer to restrain from classification and request a human expert diagnosis. The paper investigates two techniques for isolated instance identification, based on clustering and one-class support vector machines, which represent two different approaches to multidimensional outlier detection. The prediction quality for isolated instances in breast cancer image data is evaluated using the random forest algorithm and found to be substantially inferior to the prediction quality for non-isolated instances. Each of the two techniques is then used to create a novelty detection model which can be combined with a classification model and used at the time of prediction to detect instances for which the latter cannot be reliably applied. Novelty detection is demonstrated to improve random forest prediction quality and argued to deserve further investigation in medical applications.